Roses Are Red, Violets Are Blue... but Should Vqa Expect Them To?

Corentin Kervadec, Grigory Antipov, Moez Baccouche, Christian Wolf

Abstract

Models for Visual Question Answering (VQA) are notorious for their tendency to rely on dataset biases, as the large and unbalanced diversity of questions and concepts tends to prevent models from learning to "reason", leading them to perform "educated guesses" instead. In this paper, we claim that the standard evaluation metric, which consists in measuring the overall in-domain accuracy, is misleading. Since questions and concepts are unbalanced, this tends to favor models which exploit subtle training set statistics. Alternatively, naively introducing artificial distribution shifts between train and test splits is also not completely satisfying. We propose the GQA-OOD benchmark designed to overcome these concerns: we measure and compare accuracy over both rare and frequent question-answer pairs, and argue that the former is better suited to the evaluation of reasoning abilities. In a large-scale study involving 7 VQA models and 3 bias reduction techniques, we experimentally demonstrate that these models fail to address questions involving infrequent concepts and provide recommendations for future directions of research.

Introduction: The "Clever Hans" Problem in VQA

In Visual Question Answering, models often achieve high accuracy by exploiting "language priors" rather than truly understanding the visual scene. For instance, if a dataset predominantly features red roses, a model might learn to answer "red" to any question about a rose's color without ever looking at the image. These "Clever Hans" systems perform well on standard benchmarks but fail in real-world scenarios where distributions shift.

The core issue is that standard evaluation metrics—overall accuracy—are dominated by frequent, easy-to-guess cases. While previous benchmarks like VQA-CP attempted to fix this by manually swapping answer distributions, they often created "unnatural" shifts that encouraged models to solve the benchmark's specific trick rather than learn general reasoning.

The GQA-OOD Benchmark

To provide a more surgical diagnostic of model behavior, we introduce GQA-OOD. This benchmark is built upon the GQA dataset, leveraging its rich scene-graph annotations to organize questions into fine-grained "local groups" based on context (e.g., "objects on walls").

Figure 1: Illustration of the head vs. tail distribution for the question "What is on the wall?". Models often predict frequent "head" answers like "Picture" or "Mirror" instead of the rare "tail" answer "Star".

Within each group, we divide the test set into "head" samples (frequent answers in the training set) and "tail" samples (rare, out-of-distribution answers). We introduce Acc-tail (accuracy on rare samples) as the primary metric for reasoning. A model that truly reasons should perform consistently across both head and tail distributions, whereas a biased model will see a sharp drop on the tail.

Key Insights and Experimental Results

Our large-scale study involving state-of-the-art models like LXMERT, MCAN, and BAN reveals a significant "Reasoning Gap." Even the best-performing model, LXMERT, experiences a drop of roughly 10 points when moving from frequent to rare answer pairs.

Figure 5: (a) Accuracy on tail samples (Acc-tail) and (b) Confusion percentage across different models as a function of tail size.

The Failure of Bias Mitigation

Surprisingly, common bias-reduction techniques (e.g., RUBi or LM) often fail to improve Acc-tail. Instead of teaching reasoning, these methods frequently "flatten" model confidence, sometimes hurting performance on frequent cases without providing any benefit on rare ones. For example, adding the "LM" technique to a standard BUTD model caused Acc-tail to plummet from 42.1% to 32.2%.

Vision as a Bottleneck

By evaluating a "Visual Oracle" (a model with perfect ground-truth visual information), we found its performance remained nearly flat across head and tail distributions. This suggests that the primary path toward reducing language bias lies in improving visual perception and grounding.

Conclusion and Future Directions

GQA-OOD exposes that current VQA progress is heavily propped up by memorizing dataset statistics. Models excel at binary questions but rely almost entirely on bias for open-ended "query" type questions. Future research should move away from simply penalizing known biases and toward architectures that can handle open-ended queries through robust visual reasoning rather than statistical shortcuts.

Summary updated

Abstract

Models for Visual Question Answering (VQA) are notorious for their tendency to rely on dataset biases, as the large and unbalanced diversity of questions and concepts involved and tends to prevent models from learning to reason, leading them to perform educated guesses instead. In this paper, we claim that the standard evaluation metric, which consists in measuring the overall in-domain accuracy, is misleading. Since questions and concepts are unbalanced, this tends to favor models which exploit subtle training set statistics. Alternatively, naively introducing artificial distribution shifts between train and test splits is also not completely satisfying. First, the shifts do not reflect real-world tendencies, resulting in unsuitable models; second, since the shifts are handcrafted, trained models are specifically designed for this particular setting, and do not generalize to other configurations. We propose the GQA-OOD benchmark designed to overcome these concerns: we measure and compare accuracy over both rare and frequent question-answer pairs, and argue that the former is better suited to the evaluation of reasoning abilities, which we experimentally validate with models trained to more or less exploit biases. In a large-scale study involving 7 VQA models and 3 bias reduction techniques, we also experimentally demonstrate that these models fail to address questions involving infrequent concepts and provide recommendations for future directions of research.
BibTeX
@inproceedings{kervadec2021roses,
  author = {Corentin Kervadec and Grigory Antipov and Moez Baccouche and Christian Wolf},
  title = {{Roses Are Red, Violets Are Blue... but Should Vqa Expect Them To?}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2021},
  url = {https://openreview.net/forum?id=q4SCPdyewrG},
}
Figure